What is GEO? How an AI loop took a website from 35 to 81 points

GEO (Generative Engine Optimization) is the practice of optimising a website so that AI systems like ChatGPT, Claude, Perplexity or Google AI Overviews can find it, understand it and name it as a source in their answers. Felix Degeler, founder of Degeler Consulting (DC) in Vienna, built a tool with Claude that checks websites and lets AI close the gaps it finds, in a loop. On the KITALE website, the score rose from 35 to 81 out of 100 points.

I'm not a GEO expert. So I built myself a tool with Claude that automatically scans websites for their visibility to AI search engines. The first result surprised even me.

What is GEO, and how is it different from SEO?

GEO is SEO for AI answers: instead of a spot on the Google results list, the goal is to be named in an AI's answer.

SEO optimises for rankings, meaning a page appears as high as possible for a search query. GEO optimises for citations: a language model should pick up your statement, repeat it and point to your page. Many basics overlap, such as clean technology, clear page structure and content with substance.

The difference is stark: an AI answer usually names only a few sources. If you don't appear there, you simply don't exist for that search. And AI only recommends what it can find and understand.

How widespread AI search is today can only be estimated roughly. In HubSpot's State of Marketing 2026, 50 percent of surveyed end customers say they use AI-assisted search. That is a vendor survey across B2B and B2C, not a reliable figure for purchasing decisions in mid-sized companies.

How did the KITALE website go from 35 to 81 points?

In two steps: our tool found the gaps, Claude fixed them, without me deciding what needed to be done.

The tool lives in our website toolkit, fed with the collected knowledge of experienced GEO specialists. The first analysis of the newly built KITALE website came out at 35 out of 100 points.

Then I simply let Claude fix the issues the tool itself had found. The analysis finds the gaps, the AI closes them. Result afterwards: 81 out of 100 points. Why we prefer building tools like this ourselves over subscribing to them is explained in Build your own tools instead of subscriptions.

How do I make my website visible to AI search?

Eight checkpoints cover the basics: crawler access, llms.txt, structured data, dates, question-and-answer structure, quotable paragraphs, clear brand naming and sameAs links. These are exactly the points our GEO loop checks. Here is how to go about it.

First, allow AI crawlers in robots.txt: robots.txt is the file in the root of your domain that tells bots what they may read. The crawlers of the major providers, such as GPTBot, ClaudeBot or PerplexityBot, follow it. If the file locks them out, sometimes through a default setting nobody knows about, no AI can read your content, however good it is.

Second, create an llms.txt: llms.txt is a Markdown text file at yourdomain.com/llms.txt that tells language models in a few lines who you are and which pages matter most. It is a young proposal, not an established standard, and whether the major systems use it is not proven. We create one anyway, because it takes little effort and a clean summary of your own site helps regardless.

Third, structured data with Schema.org as JSON-LD: JSON-LD is a machine-readable block in the source code that records what is on the page. Types like Organization, Article, FAQPage and BreadcrumbList tell a machine explicitly whether it is looking at a company, an article or a set of questions. What is marked up does not have to be guessed from the running text.

Fourth, maintain dates: datePublished and dateModified record when content was published and when it was last revised. Search engines and AI systems need this to tell current statements from outdated ones. A page without a date looks the same to a machine as to a person: probably old.

Fifth, question-and-answer structure: headings phrased the way people ask, followed immediately by a one- or two-sentence answer. People ask an AI questions, not keywords. A page that asks the same question and answers it directly fits that query exactly. This article is built that way itself.

Sixth, quotable paragraphs: a quotable paragraph is one that is still correct and understandable when taken out of context. That means one idea per paragraph, technical terms explained on first use, no references like as mentioned above. A language model takes pieces of text, not whole pages.

Seventh, brand and entity naming: to a machine, an entity is a uniquely identifiable thing, a person, a company, a product. Write out your name, company and location in full and always the same way. What appears consistently on your website, LinkedIn profile and in directories is more reliably retrievable than a single statement, the same logic as a single point of truth.

Eighth, sameAs links: sameAs is a Schema.org property that records that different profiles belong to the same entity. In your Organization or Person markup you use it to link, for example, your LinkedIn profile. This lets a machine merge the scattered information about you instead of assuming several different people with the same name. For multilingual sites, our tool also checks hreflang, meaning which language version is intended for whom.

Why is GEO a loop and not a one-off?

Because websites, search engines and AI models change constantly, every optimisation goes out of date. The lever isn't the single run but the cycle: check, find a gap, fix it, start again.

My part in it is just kicking off the loop and reviewing the results. The AI writes the fixes, but a human still has to take responsibility for them.

Where does GEO reach its limits?

GEO makes good content findable, it doesn't make weak content better. If you have nothing of your own to say, even perfect markup won't get you cited, because a model repeats what adds something, not the average of what already exists.

There is no guarantee of being named. Which sources a model draws on is decided by the provider. The factor lists circulating on the topic are mostly hypotheses from service providers, not established findings.

AI answers fluctuate. The same question returns different sources depending on model, version and phrasing, different today than tomorrow. A single sample therefore proves neither success nor failure.

Measurement is difficult. Our score measures whether a website meets the technical and content prerequisites. The 81 points of the KITALE website say that the foundations are in place, not how often an AI actually names the site.

How do I find out how visible my website is to AI?

The quickest test: ask ChatGPT, Claude or Perplexity about your company name and about the problem you solve, and see whether and how you appear.

The principle of our loop works with any website. If you want to know how visible your own site is to AI search, talk to us. With GEO in particular, the question isn't whether AI can do the work, but whether someone keeps the loop running.